libcfs/libcfs/tracefile.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/tracefile.c
  37  *
  38  * Author: Zach Brown <zab@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42
  43 #define DEBUG_SUBSYSTEM S_LNET
  44 #define LUSTRE_TRACEFILE_PRIVATE
  45 #include "tracefile.h"
  46
  47 #include <libcfs/libcfs.h>
  48
  49 /* XXX move things up to the top, comment */
  50 union trace_data_union (*trace_data[TCD_MAX_TYPES])[NR_CPUS] __cacheline_aligned;
  51
  52 char tracefile[TRACEFILE_NAME_SIZE];
  53 long long tracefile_size = TRACEFILE_SIZE;
  54 static struct tracefiled_ctl trace_tctl;
  55 struct semaphore trace_thread_sem;
  56 static int thread_running = 0;
  57
  58 atomic_t tage_allocated = ATOMIC_INIT(0);
  59
  60 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
  61                                          struct trace_cpu_data *tcd);
  62
  63 static inline struct trace_page *tage_from_list(struct list_head *list)
  64 {
  65         return list_entry(list, struct trace_page, linkage);
  66 }
  67
  68 static struct trace_page *tage_alloc(int gfp)
  69 {
  70         cfs_page_t        *page;
  71         struct trace_page *tage;
  72
  73         /*
  74          * Don't spam console with allocation failures: they will be reported
  75          * by upper layer anyway.
  76          */
  77         gfp |= CFS_ALLOC_NOWARN;
  78         page = cfs_alloc_page(gfp);
  79         if (page == NULL)
  80                 return NULL;
  81
  82         tage = cfs_alloc(sizeof(*tage), gfp);
  83         if (tage == NULL) {
  84                 cfs_free_page(page);
  85                 return NULL;
  86         }
  87
  88         tage->page = page;
  89         atomic_inc(&tage_allocated);
  90         return tage;
  91 }
  92
  93 static void tage_free(struct trace_page *tage)
  94 {
  95         __LASSERT(tage != NULL);
  96         __LASSERT(tage->page != NULL);
  97
  98         cfs_free_page(tage->page);
  99         cfs_free(tage);
 100         atomic_dec(&tage_allocated);
 101 }
 102
 103 static void tage_to_tail(struct trace_page *tage, struct list_head *queue)
 104 {
 105         __LASSERT(tage != NULL);
 106         __LASSERT(queue != NULL);
 107
 108         list_move_tail(&tage->linkage, queue);
 109 }
 110
 111 int trace_refill_stock(struct trace_cpu_data *tcd, int gfp,
 112                        struct list_head *stock)
 113 {
 114         int i;
 115
 116         /*
 117          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 118          * from here: this will lead to infinite recursion.
 119          */
 120
 121         for (i = 0; i + tcd->tcd_cur_stock_pages < TCD_STOCK_PAGES ; ++ i) {
 122                 struct trace_page *tage;
 123
 124                 tage = tage_alloc(gfp);
 125                 if (tage == NULL)
 126                         break;
 127                 list_add_tail(&tage->linkage, stock);
 128         }
 129         return i;
 130 }
 131
 132 /* return a page that has 'len' bytes left at the end */
 133 static struct trace_page *trace_get_tage_try(struct trace_cpu_data *tcd,
 134                                              unsigned long len)
 135 {
 136         struct trace_page *tage;
 137
 138         if (tcd->tcd_cur_pages > 0) {
 139                 __LASSERT(!list_empty(&tcd->tcd_pages));
 140                 tage = tage_from_list(tcd->tcd_pages.prev);
 141                 if (tage->used + len <= CFS_PAGE_SIZE)
 142                         return tage;
 143         }
 144
 145         if (tcd->tcd_cur_pages < tcd->tcd_max_pages) {
 146                 if (tcd->tcd_cur_stock_pages > 0) {
 147                         tage = tage_from_list(tcd->tcd_stock_pages.prev);
 148                         -- tcd->tcd_cur_stock_pages;
 149                         list_del_init(&tage->linkage);
 150                 } else {
 151                         tage = tage_alloc(CFS_ALLOC_ATOMIC);
 152                         if (tage == NULL) {
 153                                 printk(KERN_WARNING
 154                                        "failure to allocate a tage (%ld)\n",
 155                                        tcd->tcd_cur_pages);
 156                                 return NULL;
 157                         }
 158                 }
 159
 160                 tage->used = 0;
 161                 tage->cpu = smp_processor_id();
 162                 tage->type = tcd->tcd_type;
 163                 list_add_tail(&tage->linkage, &tcd->tcd_pages);
 164                 tcd->tcd_cur_pages++;
 165
 166                 if (tcd->tcd_cur_pages > 8 && thread_running) {
 167                         struct tracefiled_ctl *tctl = &trace_tctl;
 168                         /*
 169                          * wake up tracefiled to process some pages.
 170                          */
 171                         cfs_waitq_signal(&tctl->tctl_waitq);
 172                 }
 173                 return tage;
 174         }
 175         return NULL;
 176 }
 177
 178 static void tcd_shrink(struct trace_cpu_data *tcd)
 179 {
 180         int pgcount = tcd->tcd_cur_pages / 10;
 181         struct page_collection pc;
 182         struct trace_page *tage;
 183         struct trace_page *tmp;
 184
 185         /*
 186          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 187          * from here: this will lead to infinite recursion.
 188          */
 189
 190         if (printk_ratelimit())
 191                 printk(KERN_WARNING "debug daemon buffer overflowed; "
 192                        "discarding 10%% of pages (%d of %ld)\n",
 193                        pgcount + 1, tcd->tcd_cur_pages);
 194
 195         CFS_INIT_LIST_HEAD(&pc.pc_pages);
 196         spin_lock_init(&pc.pc_lock);
 197
 198         cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
 199                                            struct trace_page, linkage) {
 200                 if (pgcount-- == 0)
 201                         break;
 202
 203                 list_move_tail(&tage->linkage, &pc.pc_pages);
 204                 tcd->tcd_cur_pages--;
 205         }
 206         put_pages_on_tcd_daemon_list(&pc, tcd);
 207 }
 208
 209 /* return a page that has 'len' bytes left at the end */
 210 static struct trace_page *trace_get_tage(struct trace_cpu_data *tcd,
 211                                          unsigned long len)
 212 {
 213         struct trace_page *tage;
 214
 215         /*
 216          * XXX nikita: do NOT call portals_debug_msg() (CDEBUG/ENTRY/EXIT)
 217          * from here: this will lead to infinite recursion.
 218          */
 219
 220         if (len > CFS_PAGE_SIZE) {
 221                 printk(KERN_ERR
 222                        "cowardly refusing to write %lu bytes in a page\n", len);
 223                 return NULL;
 224         }
 225
 226         tage = trace_get_tage_try(tcd, len);
 227         if (tage != NULL)
 228                 return tage;
 229         if (thread_running)
 230                 tcd_shrink(tcd);
 231         if (tcd->tcd_cur_pages > 0) {
 232                 tage = tage_from_list(tcd->tcd_pages.next);
 233                 tage->used = 0;
 234                 tage_to_tail(tage, &tcd->tcd_pages);
 235         }
 236         return tage;
 237 }
 238
 239 int libcfs_debug_vmsg2(cfs_debug_limit_state_t *cdls, int subsys, int mask,
 240                        const char *file, const char *fn, const int line,
 241                        const char *format1, va_list args,
 242                        const char *format2, ...)
 243 {
 244         struct trace_cpu_data   *tcd = NULL;
 245         struct ptldebug_header   header;
 246         struct trace_page       *tage;
 247         /* string_buf is used only if tcd != NULL, and is always set then */
 248         char                    *string_buf = NULL;
 249         char                    *debug_buf;
 250         int                      known_size;
 251         int                      needed = 85; /* average message length */
 252         int                      max_nob;
 253         va_list                  ap;
 254         int                      depth;
 255         int                      i;
 256         int                      remain;
 257
 258         if (strchr(file, '/'))
 259                 file = strrchr(file, '/') + 1;
 260
 261
 262         set_ptldebug_header(&header, subsys, mask, line, CDEBUG_STACK());
 263
 264         tcd = trace_get_tcd();
 265         if (tcd == NULL)                /* arch may not log in IRQ context */
 266                 goto console;
 267
 268         if (tcd->tcd_shutting_down) {
 269                 trace_put_tcd(tcd);
 270                 tcd = NULL;
 271                 goto console;
 272         }
 273
 274         depth = __current_nesting_level();
 275         known_size = strlen(file) + 1 + depth;
 276         if (fn)
 277                 known_size += strlen(fn) + 1;
 278
 279         if (libcfs_debug_binary)
 280                 known_size += sizeof(header);
 281
 282         /*/
 283          * '2' used because vsnprintf return real size required for output
 284          * _without_ terminating NULL.
 285          * if needed is to small for this format.
 286          */
 287         for (i = 0; i < 2; i++) {
 288                 tage = trace_get_tage(tcd, needed + known_size + 1);
 289                 if (tage == NULL) {
 290                         if (needed + known_size > CFS_PAGE_SIZE)
 291                                 mask |= D_ERROR;
 292
 293                         trace_put_tcd(tcd);
 294                         tcd = NULL;
 295                         goto console;
 296                 }
 297
 298                 string_buf = (char *)cfs_page_address(tage->page) +
 299                                         tage->used + known_size;
 300
 301                 max_nob = CFS_PAGE_SIZE - tage->used - known_size;
 302                 if (max_nob <= 0) {
 303                         printk(KERN_EMERG "negative max_nob: %i\n", max_nob);
 304                         mask |= D_ERROR;
 305                         trace_put_tcd(tcd);
 306                         tcd = NULL;
 307                         goto console;
 308                 }
 309
 310                 needed = 0;
 311                 if (format1) {
 312                         va_copy(ap, args);
 313                         needed = vsnprintf(string_buf, max_nob, format1, ap);
 314                         va_end(ap);
 315                 }
 316
 317                 if (format2) {
 318                         remain = max_nob - needed;
 319                         if (remain < 0)
 320                                 remain = 0;
 321
 322                         va_start(ap, format2);
 323                         needed += vsnprintf(string_buf + needed, remain,
 324                                             format2, ap);
 325                         va_end(ap);
 326                 }
 327
 328                 if (needed < max_nob) /* well. printing ok.. */
 329                         break;
 330         }
 331
 332         if (*(string_buf+needed-1) != '\n')
 333                 printk(KERN_INFO "format at %s:%d:%s doesn't end in newline\n",
 334                        file, line, fn);
 335
 336         header.ph_len = known_size + needed;
 337         debug_buf = (char *)cfs_page_address(tage->page) + tage->used;
 338
 339         if (libcfs_debug_binary) {
 340                 memcpy(debug_buf, &header, sizeof(header));
 341                 tage->used += sizeof(header);
 342                 debug_buf += sizeof(header);
 343         }
 344
 345         /* indent message according to the nesting level */
 346         while (depth-- > 0) {
 347                 *(debug_buf++) = '.';
 348                 ++ tage->used;
 349         }
 350
 351         strcpy(debug_buf, file);
 352         tage->used += strlen(file) + 1;
 353         debug_buf += strlen(file) + 1;
 354
 355         if (fn) {
 356                 strcpy(debug_buf, fn);
 357                 tage->used += strlen(fn) + 1;
 358                 debug_buf += strlen(fn) + 1;
 359         }
 360
 361         __LASSERT(debug_buf == string_buf);
 362
 363         tage->used += needed;
 364         __LASSERT (tage->used <= CFS_PAGE_SIZE);
 365
 366 console:
 367         if ((mask & libcfs_printk) == 0) {
 368                 /* no console output requested */
 369                 if (tcd != NULL)
 370                         trace_put_tcd(tcd);
 371                 return 1;
 372         }
 373
 374         if (cdls != NULL) {
 375                 if (libcfs_console_ratelimit &&
 376                     cdls->cdls_next != 0 &&     /* not first time ever */
 377                     !cfs_time_after(cfs_time_current(), cdls->cdls_next)) {
 378                         /* skipping a console message */
 379                         cdls->cdls_count++;
 380                         if (tcd != NULL)
 381                                 trace_put_tcd(tcd);
 382                         return 1;
 383                 }
 384
 385                 if (cfs_time_after(cfs_time_current(), cdls->cdls_next +
 386                                                        libcfs_console_max_delay
 387                                                        + cfs_time_seconds(10))) {
 388                         /* last timeout was a long time ago */
 389                         cdls->cdls_delay /= libcfs_console_backoff * 4;
 390                 } else {
 391                         cdls->cdls_delay *= libcfs_console_backoff;
 392
 393                         if (cdls->cdls_delay < libcfs_console_min_delay)
 394                                 cdls->cdls_delay = libcfs_console_min_delay;
 395                         else if (cdls->cdls_delay > libcfs_console_max_delay)
 396                                 cdls->cdls_delay = libcfs_console_max_delay;
 397                 }
 398
 399                 /* ensure cdls_next is never zero after it's been seen */
 400                 cdls->cdls_next = (cfs_time_current() + cdls->cdls_delay) | 1;
 401         }
 402
 403         if (tcd != NULL) {
 404                 print_to_console(&header, mask, string_buf, needed, file, fn);
 405                 trace_put_tcd(tcd);
 406         } else {
 407                 string_buf = trace_get_console_buffer();
 408
 409                 needed = 0;
 410                 if (format1 != NULL) {
 411                         va_copy(ap, args);
 412                         needed = vsnprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE, format1, ap);
 413                         va_end(ap);
 414                 }
 415                 if (format2 != NULL) {
 416                         remain = TRACE_CONSOLE_BUFFER_SIZE - needed;
 417                         if (remain > 0) {
 418                                 va_start(ap, format2);
 419                                 needed += vsnprintf(string_buf+needed, remain, format2, ap);
 420                                 va_end(ap);
 421                         }
 422                 }
 423                 print_to_console(&header, mask,
 424                                  string_buf, needed, file, fn);
 425
 426                 trace_put_console_buffer(string_buf);
 427         }
 428
 429         if (cdls != NULL && cdls->cdls_count != 0) {
 430                 string_buf = trace_get_console_buffer();
 431
 432                 needed = snprintf(string_buf, TRACE_CONSOLE_BUFFER_SIZE,
 433                          "Skipped %d previous similar message%s\n",
 434                          cdls->cdls_count, (cdls->cdls_count > 1) ? "s" : "");
 435
 436                 print_to_console(&header, mask,
 437                                  string_buf, needed, file, fn);
 438
 439                 trace_put_console_buffer(string_buf);
 440                 cdls->cdls_count = 0;
 441         }
 442
 443         return 0;
 444 }
 445 EXPORT_SYMBOL(libcfs_debug_vmsg2);
 446
 447 void
 448 libcfs_assertion_failed(const char *expr, const char *file,
 449                         const char *func, const int line)
 450 {
 451         libcfs_debug_msg(NULL, 0, D_EMERG, file, func, line,
 452                          "ASSERTION(%s) failed\n", expr);
 453         lbug_with_loc(file, func, line);
 454 }
 455 EXPORT_SYMBOL(libcfs_assertion_failed);
 456
 457 void
 458 trace_assertion_failed(const char *str,
 459                        const char *fn, const char *file, int line)
 460 {
 461         struct ptldebug_header hdr;
 462
 463         libcfs_panic_in_progress = 1;
 464         libcfs_catastrophe = 1;
 465         mb();
 466
 467         set_ptldebug_header(&hdr, DEBUG_SUBSYSTEM, D_EMERG, line,
 468                             CDEBUG_STACK());
 469
 470         print_to_console(&hdr, D_EMERG, str, strlen(str), file, fn);
 471
 472         LIBCFS_PANIC("Lustre debug assertion failure\n");
 473
 474         /* not reached */
 475 }
 476
 477 static void
 478 panic_collect_pages(struct page_collection *pc)
 479 {
 480         /* Do the collect_pages job on a single CPU: assumes that all other
 481          * CPUs have been stopped during a panic.  If this isn't true for some
 482          * arch, this will have to be implemented separately in each arch.  */
 483         int                    i;
 484         int                    j;
 485         struct trace_cpu_data *tcd;
 486
 487         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 488
 489         tcd_for_each(tcd, i, j) {
 490                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 491                 tcd->tcd_cur_pages = 0;
 492
 493                 if (pc->pc_want_daemon_pages) {
 494                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 495                         tcd->tcd_cur_daemon_pages = 0;
 496                 }
 497         }
 498 }
 499
 500 static void collect_pages_on_cpu(void *info)
 501 {
 502         struct trace_cpu_data *tcd;
 503         struct page_collection *pc = info;
 504         int i;
 505
 506         spin_lock(&pc->pc_lock);
 507         tcd_for_each_type_lock(tcd, i) {
 508                 list_splice_init(&tcd->tcd_pages, &pc->pc_pages);
 509                 tcd->tcd_cur_pages = 0;
 510                 if (pc->pc_want_daemon_pages) {
 511                         list_splice_init(&tcd->tcd_daemon_pages, &pc->pc_pages);
 512                         tcd->tcd_cur_daemon_pages = 0;
 513                 }
 514         }
 515         spin_unlock(&pc->pc_lock);
 516 }
 517
 518 static void collect_pages(struct page_collection *pc)
 519 {
 520         CFS_INIT_LIST_HEAD(&pc->pc_pages);
 521
 522         if (libcfs_panic_in_progress)
 523                 panic_collect_pages(pc);
 524         else
 525                 trace_call_on_all_cpus(collect_pages_on_cpu, pc);
 526 }
 527
 528 static void put_pages_back_on_cpu(void *info)
 529 {
 530         struct page_collection *pc = info;
 531         struct trace_cpu_data *tcd;
 532         struct list_head *cur_head;
 533         struct trace_page *tage;
 534         struct trace_page *tmp;
 535         int i;
 536
 537         spin_lock(&pc->pc_lock);
 538         tcd_for_each_type_lock(tcd, i) {
 539                 cur_head = tcd->tcd_pages.next;
 540
 541                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 542                                                    struct trace_page, linkage) {
 543
 544                         __LASSERT_TAGE_INVARIANT(tage);
 545
 546                         if (tage->cpu != smp_processor_id() || tage->type != i)
 547                                 continue;
 548
 549                         tage_to_tail(tage, cur_head);
 550                         tcd->tcd_cur_pages++;
 551                 }
 552         }
 553         spin_unlock(&pc->pc_lock);
 554 }
 555
 556 static void put_pages_back(struct page_collection *pc)
 557 {
 558         if (!libcfs_panic_in_progress)
 559                 trace_call_on_all_cpus(put_pages_back_on_cpu, pc);
 560 }
 561
 562 /* Add pages to a per-cpu debug daemon ringbuffer.  This buffer makes sure that
 563  * we have a good amount of data at all times for dumping during an LBUG, even
 564  * if we have been steadily writing (and otherwise discarding) pages via the
 565  * debug daemon. */
 566 static void put_pages_on_tcd_daemon_list(struct page_collection *pc,
 567                                          struct trace_cpu_data *tcd)
 568 {
 569         struct trace_page *tage;
 570         struct trace_page *tmp;
 571
 572         spin_lock(&pc->pc_lock);
 573         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc->pc_pages,
 574                                            struct trace_page, linkage) {
 575
 576                 __LASSERT_TAGE_INVARIANT(tage);
 577
 578                 if (tage->cpu != smp_processor_id() ||
 579                     tage->type != tcd->tcd_type)
 580                         continue;
 581
 582                 tage_to_tail(tage, &tcd->tcd_daemon_pages);
 583                 tcd->tcd_cur_daemon_pages++;
 584
 585                 if (tcd->tcd_cur_daemon_pages > tcd->tcd_max_pages) {
 586                         struct trace_page *victim;
 587
 588                         __LASSERT(!list_empty(&tcd->tcd_daemon_pages));
 589                         victim = tage_from_list(tcd->tcd_daemon_pages.next);
 590
 591                         __LASSERT_TAGE_INVARIANT(victim);
 592
 593                         list_del(&victim->linkage);
 594                         tage_free(victim);
 595                         tcd->tcd_cur_daemon_pages--;
 596                 }
 597         }
 598         spin_unlock(&pc->pc_lock);
 599 }
 600
 601 static void put_pages_on_daemon_list_on_cpu(void *info)
 602 {
 603         struct trace_cpu_data *tcd;
 604         int i;
 605
 606         tcd_for_each_type_lock(tcd, i)
 607                 put_pages_on_tcd_daemon_list(info, tcd);
 608 }
 609
 610 static void put_pages_on_daemon_list(struct page_collection *pc)
 611 {
 612         trace_call_on_all_cpus(put_pages_on_daemon_list_on_cpu, pc);
 613 }
 614
 615 void trace_debug_print(void)
 616 {
 617         struct page_collection pc;
 618         struct trace_page *tage;
 619         struct trace_page *tmp;
 620
 621         spin_lock_init(&pc.pc_lock);
 622
 623         pc.pc_want_daemon_pages = 1;
 624         collect_pages(&pc);
 625         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 626                                            struct trace_page, linkage) {
 627                 char *p, *file, *fn;
 628                 cfs_page_t *page;
 629
 630                 __LASSERT_TAGE_INVARIANT(tage);
 631
 632                 page = tage->page;
 633                 p = cfs_page_address(page);
 634                 while (p < ((char *)cfs_page_address(page) + tage->used)) {
 635                         struct ptldebug_header *hdr;
 636                         int len;
 637                         hdr = (void *)p;
 638                         p += sizeof(*hdr);
 639                         file = p;
 640                         p += strlen(file) + 1;
 641                         fn = p;
 642                         p += strlen(fn) + 1;
 643                         len = hdr->ph_len - (int)(p - (char *)hdr);
 644
 645                         print_to_console(hdr, D_EMERG, p, len, file, fn);
 646
 647                         p += len;
 648                 }
 649
 650                 list_del(&tage->linkage);
 651                 tage_free(tage);
 652         }
 653 }
 654
 655 int tracefile_dump_all_pages(char *filename)
 656 {
 657         struct page_collection pc;
 658         cfs_file_t *filp;
 659         struct trace_page *tage;
 660         struct trace_page *tmp;
 661         int rc;
 662
 663         CFS_DECL_MMSPACE;
 664
 665         tracefile_write_lock();
 666
 667         filp = cfs_filp_open(filename,
 668                              O_CREAT|O_EXCL|O_WRONLY|O_LARGEFILE, 0600, &rc);
 669         if (!filp) {
 670                 if (rc != -EEXIST)
 671                         printk(KERN_ERR "LustreError: can't open %s for dump: rc %d\n",
 672                                filename, rc);
 673                 goto out;
 674         }
 675
 676         spin_lock_init(&pc.pc_lock);
 677         pc.pc_want_daemon_pages = 1;
 678         collect_pages(&pc);
 679         if (list_empty(&pc.pc_pages)) {
 680                 rc = 0;
 681                 goto close;
 682         }
 683
 684         /* ok, for now, just write the pages.  in the future we'll be building
 685          * iobufs with the pages and calling generic_direct_IO */
 686         CFS_MMSPACE_OPEN;
 687         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 688                                            struct trace_page, linkage) {
 689
 690                 __LASSERT_TAGE_INVARIANT(tage);
 691
 692                 rc = cfs_filp_write(filp, cfs_page_address(tage->page),
 693                                     tage->used, cfs_filp_poff(filp));
 694                 if (rc != (int)tage->used) {
 695                         printk(KERN_WARNING "wanted to write %u but wrote "
 696                                "%d\n", tage->used, rc);
 697                         put_pages_back(&pc);
 698                         __LASSERT(list_empty(&pc.pc_pages));
 699                         break;
 700                 }
 701                 list_del(&tage->linkage);
 702                 tage_free(tage);
 703         }
 704         CFS_MMSPACE_CLOSE;
 705         rc = cfs_filp_fsync(filp);
 706         if (rc)
 707                 printk(KERN_ERR "sync returns %d\n", rc);
 708  close:
 709         cfs_filp_close(filp);
 710  out:
 711         tracefile_write_unlock();
 712         return rc;
 713 }
 714
 715 void trace_flush_pages(void)
 716 {
 717         struct page_collection pc;
 718         struct trace_page *tage;
 719         struct trace_page *tmp;
 720
 721         spin_lock_init(&pc.pc_lock);
 722
 723         pc.pc_want_daemon_pages = 1;
 724         collect_pages(&pc);
 725         cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
 726                                            struct trace_page, linkage) {
 727
 728                 __LASSERT_TAGE_INVARIANT(tage);
 729
 730                 list_del(&tage->linkage);
 731                 tage_free(tage);
 732         }
 733 }
 734
 735 int trace_copyin_string(char *knl_buffer, int knl_buffer_nob,
 736                         const char *usr_buffer, int usr_buffer_nob)
 737 {
 738         int    nob;
 739
 740         if (usr_buffer_nob > knl_buffer_nob)
 741                 return -EOVERFLOW;
 742
 743         if (copy_from_user((void *)knl_buffer,
 744                            (void *)usr_buffer, usr_buffer_nob))
 745                 return -EFAULT;
 746
 747         nob = strnlen(knl_buffer, usr_buffer_nob);
 748         while (nob-- >= 0)                      /* strip trailing whitespace */
 749                 if (!isspace(knl_buffer[nob]))
 750                         break;
 751
 752         if (nob < 0)                            /* empty string */
 753                 return -EINVAL;
 754
 755         if (nob == knl_buffer_nob)              /* no space to terminate */
 756                 return -EOVERFLOW;
 757
 758         knl_buffer[nob + 1] = 0;                /* terminate */
 759         return 0;
 760 }
 761
 762 int trace_copyout_string(char *usr_buffer, int usr_buffer_nob,
 763                          const char *knl_buffer, char *append)
 764 {
 765         /* NB if 'append' != NULL, it's a single character to append to the
 766          * copied out string - usually "\n", for /proc entries and "" (i.e. a
 767          * terminating zero byte) for sysctl entries */
 768         int   nob = strlen(knl_buffer);
 769
 770         if (nob > usr_buffer_nob)
 771                 nob = usr_buffer_nob;
 772
 773         if (copy_to_user(usr_buffer, knl_buffer, nob))
 774                 return -EFAULT;
 775
 776         if (append != NULL && nob < usr_buffer_nob) {
 777                 if (copy_to_user(usr_buffer + nob, append, 1))
 778                         return -EFAULT;
 779
 780                 nob++;
 781         }
 782
 783         return nob;
 784 }
 785
 786 int trace_allocate_string_buffer(char **str, int nob)
 787 {
 788         if (nob > 2 * CFS_PAGE_SIZE)            /* string must be "sensible" */
 789                 return -EINVAL;
 790
 791         *str = cfs_alloc(nob, CFS_ALLOC_STD | CFS_ALLOC_ZERO);
 792         if (*str == NULL)
 793                 return -ENOMEM;
 794
 795         return 0;
 796 }
 797
 798 void trace_free_string_buffer(char *str, int nob)
 799 {
 800         cfs_free(str);
 801 }
 802
 803 int trace_dump_debug_buffer_usrstr(void *usr_str, int usr_str_nob)
 804 {
 805         char         *str;
 806         int           rc;
 807
 808         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 809         if (rc != 0)
 810                 return rc;
 811
 812         rc = trace_copyin_string(str, usr_str_nob + 1,
 813                                  usr_str, usr_str_nob);
 814         if (rc != 0)
 815                 goto out;
 816
 817 #if !defined(__WINNT__)
 818         if (str[0] != '/') {
 819                 rc = -EINVAL;
 820                 goto out;
 821         }
 822 #endif
 823         rc = tracefile_dump_all_pages(str);
 824 out:
 825         trace_free_string_buffer(str, usr_str_nob + 1);
 826         return rc;
 827 }
 828
 829 int trace_daemon_command(char *str)
 830 {
 831         int       rc = 0;
 832
 833         tracefile_write_lock();
 834
 835         if (strcmp(str, "stop") == 0) {
 836                 trace_stop_thread();
 837                 memset(tracefile, 0, sizeof(tracefile));
 838
 839         } else if (strncmp(str, "size=", 5) == 0) {
 840                 tracefile_size = simple_strtoul(str + 5, NULL, 0);
 841                 if (tracefile_size < 10 || tracefile_size > 20480)
 842                         tracefile_size = TRACEFILE_SIZE;
 843                 else
 844                         tracefile_size <<= 20;
 845
 846         } else if (strlen(str) >= sizeof(tracefile)) {
 847                 rc = -ENAMETOOLONG;
 848 #ifndef __WINNT__
 849         } else if (str[0] != '/') {
 850                 rc = -EINVAL;
 851 #endif
 852         } else {
 853                 strcpy(tracefile, str);
 854
 855                 printk(KERN_INFO "Lustre: debug daemon will attempt to start writing "
 856                        "to %s (%lukB max)\n", tracefile,
 857                        (long)(tracefile_size >> 10));
 858
 859                 trace_start_thread();
 860         }
 861
 862         tracefile_write_unlock();
 863         return rc;
 864 }
 865
 866 int trace_daemon_command_usrstr(void *usr_str, int usr_str_nob)
 867 {
 868         char *str;
 869         int   rc;
 870
 871         rc = trace_allocate_string_buffer(&str, usr_str_nob + 1);
 872         if (rc != 0)
 873                 return rc;
 874
 875         rc = trace_copyin_string(str, usr_str_nob + 1,
 876                                  usr_str, usr_str_nob);
 877         if (rc == 0)
 878                 rc = trace_daemon_command(str);
 879
 880         trace_free_string_buffer(str, usr_str_nob + 1);
 881         return rc;
 882 }
 883
 884 int trace_set_debug_mb(int mb)
 885 {
 886         int i;
 887         int j;
 888         int pages;
 889         int limit = trace_max_debug_mb();
 890         struct trace_cpu_data *tcd;
 891
 892         if (mb < num_possible_cpus())
 893                 return -EINVAL;
 894
 895         if (mb > limit) {
 896                 printk(KERN_ERR "Lustre: Refusing to set debug buffer size to "
 897                        "%dMB - limit is %d\n", mb, limit);
 898                 return -EINVAL;
 899         }
 900
 901         mb /= num_possible_cpus();
 902         pages = mb << (20 - CFS_PAGE_SHIFT);
 903
 904         tracefile_write_lock();
 905
 906         tcd_for_each(tcd, i, j)
 907                 tcd->tcd_max_pages = (pages * tcd->tcd_pages_factor) / 100;
 908
 909         tracefile_write_unlock();
 910
 911         return 0;
 912 }
 913
 914 int trace_set_debug_mb_usrstr(void *usr_str, int usr_str_nob)
 915 {
 916         char     str[32];
 917         int      rc;
 918
 919         rc = trace_copyin_string(str, sizeof(str), usr_str, usr_str_nob);
 920         if (rc < 0)
 921                 return rc;
 922
 923         return trace_set_debug_mb(simple_strtoul(str, NULL, 0));
 924 }
 925
 926 int trace_get_debug_mb(void)
 927 {
 928         int i;
 929         int j;
 930         struct trace_cpu_data *tcd;
 931         int total_pages = 0;
 932
 933         tracefile_read_lock();
 934
 935         tcd_for_each(tcd, i, j)
 936                 total_pages += tcd->tcd_max_pages;
 937
 938         tracefile_read_unlock();
 939
 940         return (total_pages >> (20 - CFS_PAGE_SHIFT)) + 1;
 941 }
 942
 943 static int tracefiled(void *arg)
 944 {
 945         struct page_collection pc;
 946         struct tracefiled_ctl *tctl = arg;
 947         struct trace_page *tage;
 948         struct trace_page *tmp;
 949         struct ptldebug_header *hdr;
 950         cfs_file_t *filp;
 951         int rc;
 952
 953         CFS_DECL_MMSPACE;
 954
 955         /* we're started late enough that we pick up init's fs context */
 956         /* this is so broken in uml?  what on earth is going on? */
 957         cfs_daemonize("ktracefiled");
 958
 959         spin_lock_init(&pc.pc_lock);
 960         complete(&tctl->tctl_start);
 961
 962         while (1) {
 963                 cfs_waitlink_t __wait;
 964
 965                 cfs_waitlink_init(&__wait);
 966                 cfs_waitq_add(&tctl->tctl_waitq, &__wait);
 967                 set_current_state(TASK_INTERRUPTIBLE);
 968                 cfs_waitq_timedwait(&__wait, CFS_TASK_INTERRUPTIBLE,
 969                                     cfs_time_seconds(1));
 970                 cfs_waitq_del(&tctl->tctl_waitq, &__wait);
 971
 972                 if (atomic_read(&tctl->tctl_shutdown))
 973                         break;
 974
 975                 pc.pc_want_daemon_pages = 0;
 976                 collect_pages(&pc);
 977                 if (list_empty(&pc.pc_pages))
 978                         continue;
 979
 980                 filp = NULL;
 981                 tracefile_read_lock();
 982                 if (tracefile[0] != 0) {
 983                         filp = cfs_filp_open(tracefile,
 984                                              O_CREAT | O_RDWR | O_LARGEFILE,
 985                                              0600, &rc);
 986                         if (!(filp))
 987                                 printk(KERN_WARNING "couldn't open %s: %d\n",
 988                                        tracefile, rc);
 989                 }
 990                 tracefile_read_unlock();
 991                 if (filp == NULL) {
 992                         put_pages_on_daemon_list(&pc);
 993                         __LASSERT(list_empty(&pc.pc_pages));
 994                         continue;
 995                 }
 996
 997                 CFS_MMSPACE_OPEN;
 998
 999                 /* mark the first header, so we can sort in chunks */
1000                 tage = tage_from_list(pc.pc_pages.next);
1001                 __LASSERT_TAGE_INVARIANT(tage);
1002
1003                 hdr = cfs_page_address(tage->page);
1004                 hdr->ph_flags |= PH_FLAG_FIRST_RECORD;
1005
1006                 cfs_list_for_each_entry_safe_typed(tage, tmp, &pc.pc_pages,
1007                                                    struct trace_page, linkage) {
1008                         static loff_t f_pos;
1009
1010                         __LASSERT_TAGE_INVARIANT(tage);
1011
1012                         if (f_pos >= (off_t)tracefile_size)
1013                                 f_pos = 0;
1014                         else if (f_pos > (off_t)cfs_filp_size(filp))
1015                                 f_pos = cfs_filp_size(filp);
1016
1017                         rc = cfs_filp_write(filp, cfs_page_address(tage->page),
1018                                             tage->used, &f_pos);
1019                         if (rc != (int)tage->used) {
1020                                 printk(KERN_WARNING "wanted to write %u but "
1021                                        "wrote %d\n", tage->used, rc);
1022                                 put_pages_back(&pc);
1023                                 __LASSERT(list_empty(&pc.pc_pages));
1024                         }
1025                 }
1026                 CFS_MMSPACE_CLOSE;
1027
1028                 cfs_filp_close(filp);
1029                 put_pages_on_daemon_list(&pc);
1030                 __LASSERT(list_empty(&pc.pc_pages));
1031         }
1032         complete(&tctl->tctl_stop);
1033         return 0;
1034 }
1035
1036 int trace_start_thread(void)
1037 {
1038         struct tracefiled_ctl *tctl = &trace_tctl;
1039         int rc = 0;
1040
1041         mutex_down(&trace_thread_sem);
1042         if (thread_running)
1043                 goto out;
1044
1045         init_completion(&tctl->tctl_start);
1046         init_completion(&tctl->tctl_stop);
1047         cfs_waitq_init(&tctl->tctl_waitq);
1048         atomic_set(&tctl->tctl_shutdown, 0);
1049
1050         if (cfs_kernel_thread(tracefiled, tctl, 0) < 0) {
1051                 rc = -ECHILD;
1052                 goto out;
1053         }
1054
1055         wait_for_completion(&tctl->tctl_start);
1056         thread_running = 1;
1057 out:
1058         mutex_up(&trace_thread_sem);
1059         return rc;
1060 }
1061
1062 void trace_stop_thread(void)
1063 {
1064         struct tracefiled_ctl *tctl = &trace_tctl;
1065
1066         mutex_down(&trace_thread_sem);
1067         if (thread_running) {
1068                 printk(KERN_INFO "Lustre: shutting down debug daemon thread...\n");
1069                 atomic_set(&tctl->tctl_shutdown, 1);
1070                 wait_for_completion(&tctl->tctl_stop);
1071                 thread_running = 0;
1072         }
1073         mutex_up(&trace_thread_sem);
1074 }
1075
1076 int tracefile_init(int max_pages)
1077 {
1078         struct trace_cpu_data *tcd;
1079         int                    i;
1080         int                    j;
1081         int                    rc;
1082         int                    factor;
1083
1084         rc = tracefile_init_arch();
1085         if (rc != 0)
1086                 return rc;
1087
1088         tcd_for_each(tcd, i, j) {
1089                 /* tcd_pages_factor is initialized int tracefile_init_arch. */
1090                 factor = tcd->tcd_pages_factor;
1091                 CFS_INIT_LIST_HEAD(&tcd->tcd_pages);
1092                 CFS_INIT_LIST_HEAD(&tcd->tcd_stock_pages);
1093                 CFS_INIT_LIST_HEAD(&tcd->tcd_daemon_pages);
1094                 tcd->tcd_cur_pages = 0;
1095                 tcd->tcd_cur_stock_pages = 0;
1096                 tcd->tcd_cur_daemon_pages = 0;
1097                 tcd->tcd_max_pages = (max_pages * factor) / 100;
1098                 LASSERT(tcd->tcd_max_pages > 0);
1099                 tcd->tcd_shutting_down = 0;
1100         }
1101
1102         return 0;
1103 }
1104
1105 static void trace_cleanup_on_cpu(void *info)
1106 {
1107         struct trace_cpu_data *tcd;
1108         struct trace_page *tage;
1109         struct trace_page *tmp;
1110         int i;
1111
1112         tcd_for_each_type_lock(tcd, i) {
1113                 tcd->tcd_shutting_down = 1;
1114
1115                 cfs_list_for_each_entry_safe_typed(tage, tmp, &tcd->tcd_pages,
1116                                                    struct trace_page, linkage) {
1117                         __LASSERT_TAGE_INVARIANT(tage);
1118
1119                         list_del(&tage->linkage);
1120                         tage_free(tage);
1121                 }
1122                 tcd->tcd_cur_pages = 0;
1123         }
1124 }
1125
1126 static void trace_cleanup(void)
1127 {
1128         struct page_collection pc;
1129
1130         CFS_INIT_LIST_HEAD(&pc.pc_pages);
1131         spin_lock_init(&pc.pc_lock);
1132
1133         trace_call_on_all_cpus(trace_cleanup_on_cpu, &pc);
1134
1135         tracefile_fini_arch();
1136 }
1137
1138 void tracefile_exit(void)
1139 {
1140         trace_stop_thread();
1141         trace_cleanup();
1142 }